A table in a screenshot becomes a grid you can edit and an Excel file — without the image leaving this device.
Your files never leave your browser — all processing is local.
Drop a table screenshot here
or click to browse — you can paste a screenshot too.
The first run downloads the text model (about 15–20 MB) and the table model (about 8 MB).
Runs in your browser. Your picture is not uploaded. The first run downloads the text and table models once, and the ONNX Runtime script loads from cdn.jsdelivr.net.
Most “image to Excel” services upload the picture, and the pictures people convert are account statements, invoices and sales reports. This page keeps everything in the tab: PP-OCRv5 — the same engine as Image to Text — reads the words, and SLANet_plus, an Apache-2.0 table-structure model of about 8 MB, decides which word belongs to which cell. Both are downloaded once and cached; the OCR model is already on disk if you have used Image to Text.
Paste a screenshot or drop a photo, and the result appears as an editable grid next to the original. Fix a cell in place, then download XLSX or CSV, or copy the grid as TSV and paste it straight into Excel or Google Sheets. Numbers with thousands separators, negatives in parentheses or a percent sign are written as numbers, not text, so sums work on the other side.
OCR is not accounting. Check each figure against the image before it goes anywhere that matters, and look hardest at amounts: a lost decimal point, a comma read as a point, and 1 mistaken for 7 are the usual errors. Merged cells, diagonal headers and handwritten tables are recognised less accurately than a printed grid with clear lines.
The spreadsheet contains the numbers and text that were visible in the picture. Formulas, cell formats and hidden rows cannot be recovered from pixels, so a Total row is a typed-in number, not a SUM. Open the result in CSV Explorer to sort and filter it before it goes into a real workbook.
No. Text recognition and table-structure detection both run in this tab. The only network requests fetch the model files the first time; your image is not part of them.
PP-OCRv5 for the text — shared with Image to Text, about 15–20 MB — and SLANet_plus for the table structure, about 8 MB. Both are official PaddleOCR exports under Apache-2.0, downloaded once and cached.
Treat them as a draft. Decimal points, thousands separators and 1 versus 7 are commonly misread, so compare every amount with the image before the file is used for anything financial. The editable grid exists so you can correct a cell before export.
Merged cells, diagonal (slashed) headers, handwriting and tables without clear lines are recognised less accurately. A printed table with visible gridlines is the supported case; expect to fix more cells on anything else.
No. A picture only shows values, so every cell is written as the number or text that was visible. A total that was a formula in the original is a plain number here.
Read a paragraph of printed text from a photo, not a table.
Open the exported CSV and filter, sort or query it in the browser.
Black out private details in a screenshot before you share it.